governance-policy

Rather than directly regulating AI algorithms or models themselves, this approach uses "controlling who can access how much training compute" as a policy lever — since training frontier AI models is inseparable from large quantities of specific-spec chips, and the physical supply chain for those chips is comparatively easy to track and control, making compute itself one of the most mainstream AI governance tools in use today.

A line drawn with a concrete number — once a chip's performance or a training run's scale crosses it, reporting obligations, export restrictions, or stricter review automatically kick in, letting regulators skip evaluating how dangerous each individual model is and just check it against a threshold.